PrexSyn

Introduction

PrexSyn is an efficient, accurate, and programmable framework for synthesizable molecular design. It is based on a decoder-only transformer architecture that autoregressively generates postfix notations of synthesis¹ (a molecular representation based on chemical reactions and purchasable building blocks) conditioned on molecular descriptors.

PrexSyn is trained on a billion-scale datastream of postfix notations paired with molecular descriptors using only two GPUs and 32 CPU cores in two days. This is made possible by PrexSyn Engine, a real-time, high-throughput C++-based data generation pipeline.

Capabilities

Capability	Input	Output
Chemical space projection	Graph / SMILES
Fingerprint/descriptor based generation	Fingerprint / descriptor
Molecular sampling	Scoring functions

Performance

Capability	Result
Record-high accuracy and speed in chemical space projection and fingerprint/descriptor-based generation
Record-high sample efficiency in molecular sampling against scoring functions

Resources

Repositories

• PrexSyn: https://github.com/luost26/prexsyn
• PrexSyn Engine: The C++ backend that provides a high-throughput training data pipeline and a fast synthesis detokenizer. https://github.com/luost26/prexsyn-engine
• Data and Weights: Preprocessed chemical space data and trained model weights hosted on Hugging Face. https://huggingface.co/datasets/luost26/prexsyn-data/tree/main

Papers and Documentation

• PrexSyn Paper: Efficient and Programmable Exploration of Synthesizable Chemical Space. https://arxiv.org/abs/2512.00384
• PrexSyn Documentation: https://prexsyn.readthedocs.io

Miscellaneous

• MIT Coley Research Group: https://coley.mit.edu/

Citation

@article{luo2025prexsyn,
  title   = {Efficient and Programmable Exploration of Synthesizable Chemical Space},
  author  = {Shitong Luo and Connor W. Coley},
  year    = {2025},
  journal = {arXiv preprint arXiv: 2512.00384}
}

---

1. Projecting Molecules into Synthesizable Chemical Spaces. https://arxiv.org/abs/2406.04628 ↩